Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
University of Amsterdam From the SelectedWorks of Jeroen Hinloopen 2012 A K-sample homogeneity test: the Harmonic Weighted Mass index Jeroen Hinloopen Rien Wagenvoort Charles van Marrewijk Available at: http://works.bepress.com/hinloopen/13/ International Econometric Review (IER) A k-sample homogeneity test: the Harmonic Weighted Mass index Jeroen Hinloopen Rien J.L.M. Wagenvoort Charles van Marrewijk University of Amsterdam European Investment Bank Utrecht University ABSTRACT We propose a quantification of the p-p plot that assigns equal weight to all distances between the respective distributions: the surface between the p-p plot and the diagonal. This surface is labeled the Harmonic Weighted Mass (HWM) index. We introduce the diagonal-deviation (d-d) plot that allows the index to be computed exactly under all circumstances. This two-dimensional d-d plot accommodates a straightforward extension to the k-sample HWM index, with k > 2. A Monte Carlo simulation based on an example involving long-term sovereign credit ratings illustrates the power of the HWM test. Key words: k-sample EDF test, p-p plot, d-d plot, Power JEL Classifications: C12, C14 1. INTRODUCTION To determine whether samples are drawn from the same distribution, Empirical Distribution Function (EDF) tests can be used if the underlying population distributions are not known.1 EDF tests quantify in one way or the other percentile-percentile (p-p) plots: the scatter plot of percentiles of two distributions for all entries of their joint support.2 In this paper we introduce a new EDF statistic: the Harmonic Weighted Mass (HWM) index. It corresponds to the surface between the (k-dimensional) p-p plot and the diagonal, up to a scaling factor that depends on sample sizes. For two balanced samples without ties the HWM index closely resembles the L1–version of the Fisz-Cramér-von Mises statistic (hereafter denoted by L1–FCvM, see Schmidt and Trede, 1995). The HWM index differs from the L1–FCvM statistic when there are ties in that it is Jeroen Hinloopen, University of Amsterdam, FEB/ASE, Roetersstraat 11, 1018 WB Amsterdam, The Netherlands, (email: [email protected]). Rien J.L.M. Wagenvoort, European Investment Bank, Economic and Financial Studies, 100 Boulevard Konrad Adenauer, Luxemburg, Luxemburg; (email: [email protected]). Charles van Marrewijk, Utrecht University, Utrecht School of Economics, Kriekenpitplein 21-22, 3584 EC Utrecht, The Netherlands; (email: [email protected]). We gratefully acknowledge the comments of the Editor (Asad Zaman) and of two anonymous referees. Thanks are further due to Robert Waldmann and Sanne Zwart, to seminar participants at the University of Adelaide, the University of Amsterdam, the University of Cyprus, and the Erasmus University Rotterdam, and at the CFEERCIM 2010 conference (London) for constructive inputs. All computing codes (in GAUSS) are available upon request. The usual disclaimer applies. 1 Examples include the Kolmogorov-Smirnov (KS) test (see Kolmogorov (1933) and Smirnov, 1939), the FiszCramér-von Mises (FCvM) test (see Cramér (1928), Fisz (1960) and von Mises, 1931), the Kuiper (K) test (see Kuiper, 1960), and the Anderson-Darling (AD) test (see Anderson and Darling, 1952). 2 The obvious alternative is the scatter plot of order statistics of two distributions, the so-called quantile-quantile (q-q) plot. If distributions differ in scale and location only, q-q plots consist of straight lines, which is an attractive property when only differences in the shape of the respective distributions are of interest (Wilk and Gnanadesikan, 1968). But this is a drawback, of course, if differences in scale and location are to be revealed as well (see also Holmgren, 1995). 17 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… invariant to the position of the tie in the sequence of order statistics. This makes it a more robust statistic. Extending the HWM index to the simultaneous comparison of k > 2 samples requires first the introduction of the k-dimensional p-p plot. The diagonal of this plot corresponds to the line that cuts all 2-dimensional spaces in equal halves. The k-sample HWM index then corresponds to the surface between this ‘diagonal’ and the k-dimensional p-p plot. This surface is not uniquely defined however. An obvious norm is to consider for every point the shortest distance between the k-dimensional p-p plot and the diagonal. We derive the expression with which the HWM index can be computed exactly under all circumstances. For that we introduce the diagonal-deviation (d-d) plot: the projection of the p–p plot onto the diagonal. Being able to compute the HWM index exactly allows us to simulate its distribution in all cases for which it is not known. For two balanced samples without ties, Hinloopen and Wagenvoort (2010) derive analytically the distribution of the L1–FCvM statistic. For this special case that distribution also applies to the HWM index. Our Monte Carlo experiments show that in most cases the effect of ties on the concomitant critical values of the 2-sample HWM index is negligible, and that the sample size correction factor deals adequately with unbalanced samples. For k > 2 we provide significance tables up to k = 15. Evidently, none of the existing EDF tests dominates any of the other under all circumstances. Strictly speaking therefore, any sample comparison must involve the computation of all tests to rule out type-II errors as much as possible. Schmidt and Trede (1995) show that the L1– version of the FCvM statistic has nearly the same power as the classical L2–version of the FCvM statistic when samples are drawn from a set of widely used continuous distributions: Pareto, Log-normal, and Singh-Maddala. This result extends to the HWM index since, in these cases, ties are absent. To assess the power of the HWM test, we perform a Monte Carlo simulation based on an example involving sovereign credit ratings. It appears that the extended Anderson-Darling (AD) test (see Scholz and Stephens, 1987) has more power than the HWM test when one of the three main rating distributions (Fitch, Moody’s and Standard & Poor) is replaced by a uniform distribution. However, the HWM test can outperform AD if, in addition to the distribution homogeneity assumption, also the independency assumption is violated. Indeed, the HWM test is expected to have more power than any of the other EDF tests when the p-p plot remains ‘close’ to the diagonal over the entire probability space (while samples are drawn from different distributions) as it is the only EDF test that assigns equal weight to all distances between the respective distributions. It should therefore join the basket of EDF tests, also because in these situations a visual inspection of the p-p plot easily leads to incorrect conclusions. 2. THE 2-SAMPLE HWM INDEX Let be the set of cumulative distribution functions. The p-p plot based on F1 and F2 belonging to depicts for every domain value z from the joint support of F1 and F2 the percentiles of one distribution relative to the other: F ( z) (2.1) z 1 . F2 ( z ) 18 International Econometric Review (IER) This is a plot in the 2-dimensional simplex. An example of a p-p plot for two continuous random variables is given in Panel a of Figure 2.1. Clearly p-p plots are mappings from [0, 1] onto [0, 1] and depict the correspondence between the two underlying distributions in probability space. They are key to the hypothesis underlying EDF tests: H0: F̂1 = F̂2, where F̂ is an Empirical Distribution Function. For ease of notation, in the remainder of the paper we use Fi, i = 1, …, K to denote an empirical distribution function (rather than a cumulative distribution function). Figure 2.1 Theoretical p-p plot (panel a) and HWM index (panel b). Panel a Panel b 1 1 d F2 ( x) 0 F2 ( x) F1 ( x) 1 0 F1 ( x) 1 2.1. Definition Because the p-p plot coincides with the diagonal if, and only if, the two underlying distributions are identical, there are various well-known statistics to test H0 that are based on the distance between the p-p plot and the diagonal (see Hinloopen and Wagenvoort (2010) for an overview). In particular, the L2–version of the FCvM test sums up over all squared distances d (see Figure 2.1, Panel a) and, graphically, the Anderson-Darling test augments the FCvM-test by weighing every squared distance with the product of the distance between 0 and the centre of d, and the distance between 1 and the centre of d. We propose the area between the diagonal and the p-p plot as the criterion for validating H0 (see Panel b of Figure 2.1). Because this area reflects the extent to which the probability mass of the two underlying distributions is ‘in harmony’, we label it the Harmonic Weighted Mass (HWM) index. The HWM index requires a continuous p-p plot. But for discrete random variables the p-p plot is also discrete. The continuous analogue of a discrete p-p plot is obtained by connecting the points of the discrete p-p plot through straight lines (see Girling, 2000). Let X1 and X2 be two random variables with empirical distribution functions F1(x) and F2(x) respectively. The coordinates of the resulting piece-wise linear continuous p-p plot can be defined as: ~ F1 ( x) F1 ( zi 1 ) F1 ( zi ) ~ F ( x) F ( z ) (1 ) F ( z ) , 2 i 2 2 i 1 where z∙≡∙{z1,…,zm} are the ordered domain values of the joint support of X1 and X2, with 0∙≤∙∙≤∙1. 19 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… The 2-sample HWM index is then defined as: ~ ~ ~ Definition 2.1. HWM ( F1 , F2 ) S (n1 , n2 ) F2 ( x) F1 ( x) dF1 ( x), where the sample size correction factor S (n1 , n2 ) is taken from Rosenblatt (1952) and Fisz (1960). It reads as: n1n2 S (n1, n2 ) , (2.2) n1 n2 and speeds up convergence of the finite sample distribution towards the limiting distribution. It also makes the analytical critical percentile values of the HWM index good approximations for unbalanced samples (see Section 2.3). By virtue of the underlying p-p plot the HWM index is a non-parametric and distribution free mapping. It has two further properties (proofs of all properties are in Appendix A, Section A.1): Property 2.P1. (equality): HWM ( F1 , F2 ) 0 z a, b : F1 ( z) F2 ( z). Property 2.P2. (order irrelevance): HWM ( F1 , F2 ) HWM ( F2 , F1 ). As an illustration consider the water level of the river Meuse as it enters The Netherlands at Borgharen Dorp in 1990 and 1993 (see Hinloopen, 1997, Chapter 5). In 1993 the Southern part of Holland was plagued by severe floods and it is of interest to know whether the entire year 1993 experienced exceptionally high water levels. Table 2.1 contains the maximum water levels in millimeters for both 1990 and 1993 recorded on each last day of every month. The resulting discrete sample p-p plot, its continuous counterpart, and the concomitant HWM index are all depicted in Figure 2.2. Month 1990 1993 January 4070 4114 February 4204 3944 March 3885 3814 April 3866 3813 May 3808 3836 June 3854 3757 July 3762 3824 August 3786 3751 September 3764 3809 October 3872 3896 November 3926 3818 December 4292 4406 Table 2.1 Maximum water levels of the river Meuse at Borgharen Dorp, in millimetres, measured on every last day of every month. Source: Koninklijk Nederlands Meteorologisch Instituut, personal correspondence (see Hinloopen, 1997, Chapter 5). 20 International Econometric Review (IER) Figure 2.2 Discrete p-p plot for water levels of the river Meuse at Borgharen Dorp (The Netherlands), 1990 versus 1993 (Panel a), the corresponding continuous p-p plot (Panel b), the concomitant HWM index (Panel c), and the resulting d-d plot (Panel d). Panel a Panel b 1 1 P X 2 zi P X 2 zi 0 P X 1 zi 1 0 Panel c P X 1 zi 1 Panel d 1 1 2 di P X 2 zi 1 18 0 P X 1 zi 1 0 pi 1 2.2. Computation Note that the value of the shaded area in Panel c of Figure 2.2, which equals 7/12, is straightforward to obtain. For samples with ties this is not necessarily the case. Within-sample ties affect the number of coordinates that make up the p-p plot. As such they have no effect on the possible slope of the p-p plot. Between-sample ties may induce the p-p plot to have linear pieces with any positive slope. This complicates the exact computation of the HWM index. For computing the HWM index we introduce the diagonal-deviation plot: the projection of the p-p plot onto the diagonal.3 This means that each point (F1(z), F2(z)) on the p-p plot is projected on the average probability p(z)∙≡∙(F1(z)∙+∙F2(z))/2. Because the length of the projection vector equals d ( z ) F1 ( z) F2 ( z) 2 , we arrive at the following definition: Definition 2.2. The diagonal-deviation plot is the projection of the p-p plot onto the diagonal: p( z ) z , p( z ) F1 ( z ) F2 ( z ) / 2 , d ( z ) F1 ( z ) F2 ( z ) d ( z) 3 2. Use of this projection is not necessary for computing the exact value of the 2-sample HWM index but it allows for a straightforward extension of the HWM index to the simultaneous comparison of k > 2 samples (see Section 3). 21 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… The d-d plot is obtained by first projecting the coordinates of the discrete p-p plot onto the diagonal. This gives the coordinates of the d-d plot: D∙≡∙(p(zi), d(zi)), i = 1, 2, ..., m, where m∙≤∙n1 + n2, and (p(0), d(0))∙≡ (0,0). Next, linearly interpolate between these coordinates to obtain the d-d plot. The value of the HWM index then corresponds to the surface between the d-d plot line and the horizontal zero axis multiplied by the projection scaling factor PS(2)∙=∙√2̄ (see Panel d in Figure 2.2). Computation of the HWM index requires the exact location of all intersections of the underlying p-p plot with the diagonal. Diagonal cutting points are all points (F1(z), F2(z)) for which d(z) = 0. Their location follows from the following lemma (the proof of which is in Appendix A, Section 6.2): Lemma 2.1. When the p-p plot cuts the diagonal between points z i and z i 1 , zi z, then ( F1,n1 ( z i 1 ) F1,n1 ( z i ))( F1,n1 ( z i ) F2,n2 ( z i )) p h F1,n1 ( z i ) ( F2,n2 ( z i 1 ) F2,n2 ( z i )) ( F1,n1 ( z i 1 ) F1,n1 ( z i )) if at probability p h we have that F1, n1 ( zi ) F2, n2 ( zi ) and F1, n1 ( zi 1 ) F2, n2 ( zi 1 ), that is, the diagonal is “cut from below”, and ( F2,n2 ( z i 1 ) F2,n2 ( z i ))( F2,n2 ( z i ) F1,n1 ( z i )) p h F2,n2 ( z i ) ( F1,n1 ( z i 1 ) F1,n1 ( z i )) ( F2,n2 ( z i 1 ) F2,n2 ( z i )) if at probability p h we have that F1, n1 ( zi ) F2, n2 ( zi ) and F1, n1 ( zi 1 ) F2, n2 ( zi 1 ), that is, the diagonal is “cut from above”. Obviously, the p-p plot can cut the diagonal at points that are not in D. Let L be the set with these diagonal cutting points consisting of l entries and let I∙=∙D∙∙L∙≡∙(p*i,∙d*i∙), i∙=∙1,∙2,∙...,∙m∙+∙l, be ordered on p*i. The HWM index can then be calculated exactly under all circumstances (the proofs of all propositions are in Appendix A, Section 6.3): Proposition 2.1. m l HWM ( F1, n1 , F2, n2 ) PS (2)S (n1, n2 ) i 1 p * i p i*1 2 max{d i* , d i*1 } d i* d i*1 . 2 2.3. Hypothesis testing Being able to compute the value of the HWM index exactly allows for the simulation of significance tables under varying circumstances. For a special case however the exact finite sample distribution of the HWM index is known. 2.3.1. Finite sample distribution Consider the set of continuous distribution functions . Observe that . Also, because functions belonging to are strictly increasing on their support, there are no mass points. That is, there are no ties. Indeed, the HWM index coincides with the L1–version of the FiszCramér-von Mises statistic (Schmidt and Trede, 1995) for distributions belonging to and for two samples of equal size n. We thus introduce: Assumption 2.A1. F1 , F2 2 , 22 International Econometric Review (IER) Assumption 2.A2. n1 n2 n. Proposition 2.2. Under A1 – A2, the HWM index coincides with the L1–FCvM statistic: n1n2 1 / 2 n1 n2 F ( z ) F ( z ) . 1, n1 i 2, n 2 i n1 n2 3 / 2 i 1 Hinloopen and Wagenvoort (2010) derive the analytical solution for the probability density function of the L1–FCvM statistic under H0 and A1 – A2. For these cases, the exact critical percentiles of the HWM index are thus known. Indeed, the diagonal entries of Tables B.5 through B.8 in Appendix B correspond to these percentiles for two balanced samples up to n∙=∙20. High values of the HWM index imply a low probability that the underlying samples are drawn from the same distribution. For instance, for the data in Table 2.1 the HWM index value is 0.2381, which is substantially smaller than the critical percentiles at common significance levels. Although the southern part of The Netherlands was flooded in 1993, considering the entire year shows that the monthly maximum water levels of the river Meuse at Borgharen Dorp in 1993 were not significantly different from the levels in 1990. 2.3.2. Limiting distribution For balanced samples without ties Schmid and Trede (1995) note that the limiting distribution of the L1–FCvM statistic corresponds to the limiting distribution of the L1–norm of a Brownian bridge. Proposition 2.2 implies that the same holds for the HWM index. The analytical expression for the L1–norm of a Brownian bridge is derived by Johnson and Killeen (1983). They also tabulate the concomitant critical values. It appears that the critical values of the HWM index converge rapidly to their limiting values (see Table 1 in Hinloopen and Wagenvoort, 2010). Indeed, although for finite samples the critical percentiles of the limiting distribution differ from the exact values, this will not lead to differences in the rejection of the underlying hypothesis (see Hinloopen and Wagenvoort, 2010). 2.3.3. Ties Ties can be present in any sample, even for continuous population distributions due to rounding. Within-sample ties reduce the number of coordinates that constitute the p-p plot. Between-sample ties allow the p-p plot to remain closer to the diagonal. Ties therefore affect every EDF that quantifies a p-p plot. One way of dealing with ties is to use a randomized tiebreaking procedure (Dufour (1995), Dufour and Kiviet, 1998). Let Ui, i∙=∙1,∙...,∙n1∙+∙n2 be a random sample of n1∙+∙n2 observations from a uniform continuous distribution. The observations Z∙=∙X1∙∙ X2 can then be arranged following the order: (Z i ,U i ) (Z j ,U j ) Z i Z j or (Z i Z j and U i U j ), (2.3) which results in n1∙+∙n2 different order statistics. The test statistic is then computed for these n1∙+∙n2 different order statistics rather than the q < n1∙+∙n2 order statistics from the original samples. Alternatively the q order statistics are used and the critical values in Appendix B are applied at the possible cost of a small size distortion in the critical area. Indeed, within-sample ties possibly increase the value of the HWM index whereas between-sample ties possibly reduce 23 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… it. To assess the impact of ties on the distribution of the HWM index we have to rely on numerical simulations because assumption 2.A1 is violated. percentile 90 95 97.5 99 {0, 1} -0.047 -0.027 -0.018 -0.006 {0, 1, 2} -0.040 -0.022 -0.014 -0.006 {0, 1, 2, 3} -0.023 -0.013 -0.007 -0.002 {0,…, 4} -0.023 -0.012 -0.009 -0.003 {0,…, 5} -0.018 -0.011 -0.005 -0.003 {0,…, 6} -0.016 -0.009 -0.002 0.001 {0,…, 7} -0.012 -0.007 -0.004 -0.002 {0,…, 8} -0.011 -0.006 -0.003 -0.001 {0,…, 9} -0.012 -0.007 -0.004 -0.001 {0,…, 10} -0.008 -0.005 -0.003 -0.002 {0,…, 11} -0.010 -0.008 -0.004 -0.001 {0,…, 12} -0.006 -0.004 -0.004 -0.001 {0,…, 13} -0.004 -0.006 -0.003 -0.002 {0,…, 14} -0.005 -0.005 -0.003 -0.001 domain {0,…, 15} -0.008 -0.006 -0.006 -0.002 Table 3.2 Size distortions due to ties in balanced samples. Notes: All samples consist of 50 entries which are drawn from the integer domains in the first column, where all entries have equal probability. The resulting simulated distribution for each row consists of 10,000 independent HWM index values. As the influence of ties on the distribution of the HWM index turns out to be negligible in most cases, we report only the simulation results for an extreme situation in that all observations constitute a tie. Table 3.2 lists the size distortions of the exact critical percentiles derived by Hinloopen and Wagenvoort (2010). This distortion is defined as prob[HWM∙>∙cv]∙–∙cp, where cv is the critical value of the HWM index under A1 – A2 and cp is the concomitant probability (i.e. 10%, 5%, 2.5% and 1%). Clearly, what matters is the number of classes that underlie the sample EDFs. For instance, if there are only 2 classes, 0 and 1 say, the 90th percentile is 0.400 while without ties it equals 0.500, yielding a size distortion of -0.047. But size distortions fall rapidly when the number of classes increases. With as little as 7 classes the simulated 90th percentile is already 0.485 and the size distortion is one percentage point only. It thus seems that for the HWM index the effect of withinsample ties and between-sample ties cancel out when the number of classes is sufficiently high. Because many applications will involve less than 100% ties (the only other study we know of that considers the effect of ties, Scholz and Stephens (1987), reports simulations up to situations where 60% of all observations constitute a tie), we conjecture that the critical values shown in Appendix B are accurate under most circumstances where assumption 2.A1 is violated. Perhaps more importantly, the HWM index is less sensitive to ties than the L1–FCvM statistic. This is because the HWM index is invariant to the position of a tie whereas the FCvM statistic is not. For instance, let X1={1, 2, 3}, and X2= {1.5, 1.5, 4} with respective discrete EDFs F1,3 and F2,3. If the entries in X2 are rounded upwards, yielding X 2 {2, 2, 4} and F2, 3 , a 24 International Econometric Review (IER) between-sample tie arises at the second entry of both samples. This gives HWM ( F1,3 , F2, 3 ) = 0.2041 and L1 FCvM ( F1,3 , F2, 3 ) 0.1361. Alternatively the entries are rounded downwards, giving us X 2 {1, 1, 4} and F2,3 . A between-sample tie now occurs at the first entry of both samples. This leaves the HWM index unaffected, HWM ( F1,3 , F2,3 ) = 0.2041 while it does influence the FCvM statistic: L1 FCvM ( F1,3 , F2,3 ) 0.2722. That is, the HWM index takes the same value, whether the between-sample tie occurs at the first or second entry. The HWM index is unaffected because it is based on a continuous p-p plot where mass points are removed by spreading probability uniformly between two order statistics. In contrast, the FCvM statistic changes value depending on the position of the tie as the FCvM statistic is based on the discrete p-p plot. This makes the FCvM statistic a less robust statistic since the augmented samples X 2 and X 2 are ‘equally close’ to sample X 1 . 2.3.4. Unbalanced samples To assess the influence of unbalanced samples we again have to revert to numerical simulations. Tables B.5 through B.8 in Appendix B (Section 7.1) list the simulated percentiles of the HWM index under H0 for all possible unbalanced samples up to n = 20. Clearly, sample size correction factor (2) adequately deals with unbalanced samples. For larger samples the simulated percentiles suggest a simple rule of thumb: in case of unbalanced samples that are not in Appendix B use the analytical value of the HWM index for the largest sample size. Even in an extreme case where one sample consists of 3 entries while the other has 20 entries this yields a small approximation error. 3. THE K-SAMPLE HWM INDEX For extending the HWM index to the simultaneous comparison of k > 2 samples we first have to introduce the k-sample p-p plot: Definition 3.3. The k-sample p-p plot depicts for every domain value z from the joint support of F1 ,..., Fk the percentiles of one distribution relative to the others: F1 ( z ) z ... . Fk ( z ) Obviously k-sample p-p plots are mappings from [0, 1] to [0, 1]k-1. Hypothesis H0 then extends to: H*0: F1=…=Fk. And the generalized assumptions 2.A1 and 2.A2 respectively read as: Assumption 3.A1*. F1 ,..., Fk 2 , Assumption 3.A2*. n1 ... nk n. 3.1. Definition The k-dimensional HWM index is defined as the surface between the k-dimensional p-p plot and the ‘diagonal’ that cuts all 2-dimensional spaces in equal halves. The coordinates of this 25 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… diagonal are given by the k-dimensional vector ( ~ p ( x) ,…, ~ p ( x) ), where ~ p ( x) is the average probability at x. For k > 2 this surface is not uniquely defined however. It depends on the point of the diagonal on which probabilities are projected. An obvious candidate is to take the shortest distance between the p-p plot and the diagonal, which implies that each point of the p-p plot is projected on the concomitant average probability.4 For characterizing this distance we use the Mahalanobis distance: Definition 3.4. ~ j 1 F j ( x) p ( x) HWM ( F1 , F2 ,..., Fk ) PS (k )S (n1 , ..., nk ) k ~ 2 d~ p ( x), ~ where F j ( x) is the continuous analogue of the possibly discrete EDF of X j , ~ p ( x) 1 k ~ F j ( x) is the average probability at x, where S (n1 ,..., nk ) is a multi-sample scaling k j 1 factor, and where PS(k) is a factor that scales the projection. In particular we generalize (2) to: S (n1 ,..., n k ) 1 k j 1n j k 1 k 2 j 1 n j (3.4) . The projection scaling factor PS(k) maintains the correspondence between the surface below the d-d plot and the surface between the k-dimensional p-p plot and the diagonal. It equals the length of the diagonal of the k-dimensional p-p plot: (3.5) PS (k ) k . Note that the 2-sample version of the HWM index in Definition 2.1 corresponds to Definition 3.4 with k = 2. The properties of the 2-sample HWM index carry over to the k-sample version: Property 3.P1* (equality): HWM ( F1 , ..., Fk ) 0 z a, b: F1 ( z) ... Fk ( z). Property 3.P2* (order irrelevance): HWM ( F1 , ..., Fk ) HWM (G( F1 , ..., Fk )), where G(∙) is any perturbation of the order of its entries. 3.2. Computation Computing the k-sample HWM index involves again the construction of the underlying d-d plot. In case the p-p plot is projected on the average probabilities, the coordinates of the corresponding d-d plot for any z i z, are: p( z i ) d ( zi ) 1 k k F j ( z i ), (3.6) j 1 j 1 F j ( z i ) p( z i ) 2 k 4 , (3.7) We also considered alternatives, such as the minimum probability. This led to much more variability in the HWM index which, we think, makes it less suitable for hypotheses testing. 26 International Econometric Review (IER) i 1, ..., m, m kj 1 n j , and ( p(0), d (0)) (0,0). Obviously, Proposition 2.1 applies such that the k-sample HWM index can be computed as follows: ml p * pi*1 2 max{ d * , d * } d * d * HWM ( F1,n1 , ..., Fk ,nk ) PS (k )S (n1 , ..., nk ) i i i 1 i i 1 2 i 1 in the notation of Proposition 2.1. , 3.3. Hypotheses testing An exact formulation is not known for the finite sample distribution of any of the existing ksample EDF tests with k > 2 (Kiefer, 1959). However, as the HWM index is distribution free under assumptions 3.A1* – 3.A2*, we can simulate its distribution. The concomitant critical percentiles are in Tables B.9 through B.12 in Appendix B for k = 3, ..., 15 for sample sizes up to the point of convergence. 4. LONG-TERM SOVEREIGN DEBT CREDIT RATINGS The performance of the HWM test should be measured against its ability to discriminate between samples that are not drawn from the same population distribution. As none of the existing EDF tests dominates all other tests under all circumstances, in principle all tests must be considered at all times. However, in some cases a particular test is known to have the best power. For instance, when samples are drawn from distributions with different tails, the Anderson-Darling (AD) test outperforms all alternatives as it puts more weight on extremities. The HWM test on the other hand is expected to outperform all other tests when the p-p plot remains relatively close to the diagonal because it is the only EDF test that assigns equal weight to all distances between the respective distributions. Rating agency Credit ratings AAA, AA+ AA, AAA+, A A-, BBB+ BBB, BBBBB+, BB BB-, B B-, CCC+ CCC, CCCCC+, CC Fitch Moody’s S&P 25 6 11 5 26 10 9 6 1 0 23 13 9 10 16 13 13 4 0 1 23 13 8 8 21 13 11 4 0 1 Value test statistic 0.163 -1.690 HWM AD n1 n2 n3 80 Number of observations Table 4.3 Share (in %) of countries ordered according to their long-term sovereign debt credit ratings, September 7, 2011. Notes: The upper 90th critical percentile is 0.701 and 1.309 for HWM and AD respectively. As an illustration consider the long-term sovereign debt credit ratings of the main rating agencies Fitch, Moody’s, and Standard & Poor’s (S&P). On 7 Sept. 2011, in total 80 countries were rated by each agency (IMF, 2011). Out of 22 possible ratings, each of the 20 ratings 27 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… from AAA to CC was given to at least one country. Table 4.3 contains the rating distributions, where each ‘bin’ consists of two ratings. For example, Fitch rated 25% of the 80 countries AAA or AA+. Obviously, there are many within and between sample ties across the rating distributions. Both the HWM and the Anderson-Darling test do not reject the null-hypothesis that the ratings are drawn from the same underlying distribution at the 10 percent or higher confidence level; there are no differences in outcome of the rating methodologies across the three rating agencies. In order to compare the power of HWM and AD, we first use the rating distributions as the basis for a Monte Carlo simulation. For that we randomly draw sub-samples of 10 or 50 observations from the Moody’s and S&P rating data. A third sub-sample is drawn from the uniform distribution with integer domain [1, 2, …, 20] rather than from the Fitch rating distribution. Next we compute both HWM and AD and determine the fraction of rejections in 10,000 trials. The results are in Table 4.4. Number of observations Test n1 n2 n3 HWM AD Share of rejections at the 10% confidence level 28 39 10 83 95 50 Share of rejections at the 5% confidence level 17 28 10 76 92 50 Share of rejections at the 2.5% confidence level 11 19 10 66 87 50 Share of rejections at the 1% confidence level 5 11 10 54 78 50 Table 4.4 Power assessment in a Monte Carlo experiment. Notes: Observations are randomly drawn from the Moody’s and S&P rating distributions. A third sub- sample is drawn from the uniform distribution on the integer domain [1, 2,…,20]. Results are based on 10,000 trials. AD has more power than HWM to discriminate between the three sub-samples. In small samples of 10 observations, AD rejects H0 in nearly twice as many cases as HWM. Differences in power are smaller for larger samples. For example, at the 10% confidence level HWM rejects H0 in 83% of the cases with sub-samples of 50 observations compared to a rejection rate of 95% for AD. For the simulation underlying Table 4.4, observations are drawn independently from each distribution. However, if observations are not drawn independently, HWM can have more power than AD. For instance, if we fix the relative frequencies of the 10 bins in Table 4.3 such that distributions differ significantly, but relatively little at the extremities, examples can be found where HWM outperforms AD. Figure 4.3 depicts such a situation. It depicts fictitious rating distributions for three fictitious rating agencies.5 If these distributions are used when drawing samples of 100 observations, a different picture arises: at the 5% confidence level, HWM rejects H0 in 42.9% of all cases, whereas AD rejects H0 in 38.6% only. 5 In particular, the probabilities over the respective 10 bins for the three fictitious rating agencies are (in percentages): {5, 10, 10, 10, 15, 15, 10, 10, 10, 5}, {5, 10, 10, 18, 18, 11, 12, 6, 5, 5} and {5, 5, 6, 12, 11, 18, 18, 10, 10, 5}. 28 International Econometric Review (IER) Figure 4.3 Fictitious probabilities with which samples are drawn from the 10 rating bins for three fictitious rating agencies, yielding HMW to have higher power than AD at the 5% confidence level. 20% 16% 12% 8% 4% 0% AAA, AA+ AA, AA- A+, A A-, BBB+ Agency A BBB, BB+, BB BBBAgency B BB-, B B-, CCC+ CCC, CCC- CC+, CC Agency C 5. Discussion and conclusions The quantification of the p-p plot we propose assigns equal weight to all distances between the respective distributions. For two balanced samples without ties it is shown that the HWM index coincides with the L1–FCvM statistic (see Schmidt and Trede (1995); see also Hinloopen and Wagenvoort, 2010). The HWM index differs from the L1–FCvM statistic when there are ties in that it is invariant to the position of the tie in the sequence of order statistics. This makes it a more robust statistic. We further show how the HWM index can be extended to a homogeneity test involving k > 2 samples. For that we have introduced the k-sample pp-plot. We also show how the k-sample HWM index can be computed exactly under all circumstances. In so doing we have introduced the d-d plot: the projection of the p-p plot onto the diagonal. Both ties and unbalanced samples appear to affect the distribution of the 2-sample HWM index only mildly. However, ties can have a non-negligible effect on the distribution of the ksample HWM index for k > 2, even if the underlying population distributions consist of many classes. This is because the probability of a between-sample tie increases with k relative to the probability of a within-sample tie, and different ties have an opposite effect on the value of the HWM index. In sum, as none of the existing EDF tests, including the HWM test, outperforms all other tests under all circumstances, they should all be considered simultaneously so as to avoid type-II errors as much as possible. 29 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… APPENDIX Appendix A Proofs A.1. Proofs of properties ~ Proof of Property P1*. Let [a,b] be the common support of F j ( z ), j = 1, …, k. ~ F j ( z) of the continuity of both and HWM ( F1 , F2 ,..., Fk ) 0 because 2 ~ ~ ~ ~ p ( z ) : kj 1 F j ( z ) ~p ( z ) 0 z a, b F1 ( z) ~p, ..., Fk ( z) ~p, z a, b F1 ( z) ... Fk ( z), z a, b. QED Proof of Property P2*. HWM ( F1 , F2 ,..., Fk ) PS (k )S (n1 , ..., nk ) ~ j 1 F j ( z ) p ( z ) k ~ 2 d~ p ( z ) = PS (k ) S (n1 , ..., nk ) 2 ~ G kj 1 F j ( z ) ~ p ( z ) d~ p( z) HWM (G( F1 , ..., Fk )), where G () is any perturbation of the order of its entries. QED A.2. Proofs of Lemma 2.1 A p-p plot that cuts the diagonal from below at probability ph between observations z i and z i 1 (i.e. F1,n1 ( z i ) F2,n2 ( zi ) and F1, n1 ( zi 1 ) F2, n2 ( zi 1 )) is illustrated in Figure A.4. The diagonal runs from ( F1,n1 ( z i ), F1,n1 ( zi )) to ( F1,n1 ( z i 1 ), F1, n1 ( zi 1 )) while the p-p plot connects ( F1,n1 ( z i ), F2,n1 ( z i )) with ( F1, n1 ( zi 1 ), F2, n2 ( z i 1 )). Let a ph F1,n1 ( zi ), b F1,n1 ( zi ) F2,n2 ( zi ), c F1,n1 ( zi 1 ) F1,n1 ( zi ), and d = F2, n2 ( zi 1 ) F2, n2 ( zi ). From TAN(t) = d / c = (a + b)/a we obtain that a = bc/(d − c). Hence, ph follows. A similar argument applies in case the diagonal is cut from above. QED A.3. Proofs of propositions Proof of Proposition 2.1. The surface between the p-p plot and the diagonal equals the surface below the d-d plot multiplied by k . The surface below the d-d plot between ( pi*1 , di*1 ) and ( pi* , d i* ) equals ( pi* pi*1 )di*1 ( pi* pi*1 )(di* di*1 ) / 2 if di* di*1 , and ( pi* pi*1 )di* ( pi* pi*1 )(di*1 di* ) / 2 if di* di*1. The proposition then follows. QED Proof of Proposition 2.2. Under A1 – A2, the contribution of any two points z i and z i 1 to the L1 FCvM statistic equals F2 ( z i ) F1 ( z i ) F2 ( z i 1 ) F1 ( z i 1 ) . 2 2n The surface between the diagonal and the p-p plot from point z i to z i 1 multiplied by the scaling factor S (n1 , n2 ) n 2n equals 30 International Econometric Review (IER) n F2 ( z i 1 ) F2 ( z i ) F2 ( z i ) F1 ( z i 1 ) F1 ( z i 1 ) F1 ( z i ) 2n n 2n 2n F2 ( z i ) F1 ( z i ) F2 ( z i 1 ) F1 ( z i 1 ) , 2 2n when F2 ( z i ) F1 ( z i ) and F2 ( z i 1 ) F1 ( z i 1 ), i.e. the p-p plot is above the diagonal, and equals n F2 ( z i 1 ) F2 ( z i ) F1 ( z i ) F2 ( z i 1 ) F1 ( z i 1 ) F1 ( z i ) 2n n 2n 2n F1 ( z i ) F2 ( z i ) F1 ( z i 1 ) F2 ( z i 1 ) , 2 2n when F2 ( z i ) F1 ( z i ) and F2 ( z i 1 ) F1 ( z i 1 ), i.e. the p-p plot is below the diagonal. The contribution of any two points z i and z i 1 to the value of the HWM index is thus identical to their contribution to the value of the L1 FCvM statistic. QED Figure A.4 A p-p plot cutting the diagonal “from below” at point ( ph , ph ). F2,n2 ( zi 1 ) p-p plot F1,n1 ( zi 1 ) diagonal ph , ph d a c F1,n1 ( zi ) b t F2,n2 ( zi ) F1,n1 ( zi ) F1,n1 ( zi 1 ) 31 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… Appendix B Critical percentiles B.1. k = 2, unbalanced samples, no ties n 3 4 5 6 7 8 9 10 11 3 0.612 4 0.546 0.530 5 0.502 0.522 0.538 6 0.511 0.516 0.495 0.529 7 0.518 0.513 0.512 0.514 0.515 8 0.492 0.510 0.510 0.501 0.518 0.500 9 0.528 0.508 0.498 0.498 0.501 0.505 0.511 10 0.516 0.507 0.511 0.506 0.505 0.502 0.508 0.514 11 0.501 0.506 0.504 0.508 0.512 0.509 0.506 0.502 0.494 12 0.516 0.505 0.501 0.500 0.508 0.506 0.504 0.506 0.502 13 0.503 0.505 0.512 0.498 0.504 0.502 0.503 0.512 0.504 14 0.495 0.504 0.498 0.505 0.496 0.504 0.508 0.513 0.504 15 0.509 0.503 0.512 0.506 0.502 0.505 0.502 0.504 0.498 16 0.497 0.503 0.508 0.500 0.504 0.505 0.500 0.500 0.505 17 0.507 0.501 0.501 0.510 0.508 0.508 0.508 0.502 0.512 18 0.510 0.503 0.506 0.511 0.506 0.498 0.507 0.507 0.505 19 0.499 0.503 0.508 0.506 0.500 0.502 0.499 0.499 0.506 20 0.501 0.502 0.505 0.507 0.510 0.509 0.501 0.503 0.513 n 12 13 14 15 16 17 18 19 20 12 0.510 13 0.500 0.505 14 0.508 0.509 0.499 15 0.502 0.508 0.510 0.505 16 0.506 0.506 0.501 0.500 0.508 17 0.502 0.502 0.501 0.504 0.507 0.499 18 0.505 0.498 0.501 0.508 0.495 0.502 0.500 19 0.499 0.506 0.506 0.503 0.506 0.504 0.502 0.499 20 0.505 0.502 0.502 0.505 0.511 0.500 0.495 0.509 0.506 Table B.5 Simulated critical values at percentile 90 for unbalanced samples without ties. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. The joint number of sample entries must be at least 6 for percentile 90 to be defined properly. 32 International Econometric Review (IER) n 3 4 5 6 7 8 9 10 11 4 0.655 0.619 5 0.593 0.596 0.601 6 0.550 0.581 0.606 0.577 7 0.587 0.570 0.581 0.599 0.592 8 0.577 0.587 0.596 0.582 0.587 0.594 9 0.583 0.601 0.578 0.586 0.583 0.575 0.589 10 0.608 0.592 0.584 0.581 0.583 0.580 0.595 0.581 11 0.582 0.584 0.590 0.584 0.593 0.587 0.578 0.583 0.591 12 0.570 0.577 0.595 0.583 0.584 0.593 0.588 0.585 0.590 13 0.580 0.582 0.599 0.572 0.578 0.578 0.581 0.588 0.586 14 0.569 0.576 0.576 0.580 0.584 0.584 0.594 0.587 0.596 15 0.580 0.565 0.590 0.587 0.582 0.590 0.589 0.588 0.578 16 0.569 0.587 0.586 0.588 0.591 0.586 0.583 0.587 0.580 17 0.579 0.582 0.573 0.585 0.589 0.593 0.587 0.587 0.592 18 0.584 0.578 0.593 0.589 0.588 0.583 0.590 0.585 0.587 19 0.583 0.579 0.577 0.581 0.576 0.590 0.579 0.579 0.582 20 0.569 0.589 0.585 0.579 0.591 0.589 0.584 0.587 0.593 n 12 13 14 15 16 17 18 19 20 12 0.578 13 0.591 0.581 14 0.590 0.586 0.594 15 0.591 0.594 0.589 0.590 16 0.586 0.592 0.586 0.591 0.586 17 0.585 0.575 0.582 0.592 0.585 0.580 18 0.584 0.577 0.581 0.583 0.576 0.580 0.593 19 0.583 0.587 0.587 0.585 0.591 0.584 0.580 0.585 20 0.586 0.583 0.583 0.586 0.589 0.583 0.581 0.598 0.585 Table B.6 Simulated critical values at percentile 95 for unbalanced samples without ties. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. The joint number of sample entries must be at least 7 for percentile 95 to be defined properly. 33 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… n 3 4 5 6 7 8 9 10 11 5 0.685 0.671 0.664 6 0.629 0.645 0.661 0.674 7 0.656 0.627 0.659 0.642 0.668 8 0.615 0.663 0.658 0.656 0.656 0.656 9 0.639 0.647 0.657 0.650 0.646 0.657 0.642 10 0.658 0.651 0.657 0.645 0.667 0.659 0.653 0.648 11 0.641 0.662 0.657 0.657 0.673 0.660 0.663 0.645 0.649 12 0.645 0.650 0.658 0.653 0.651 0.662 0.658 0.662 0.655 13 0.661 0.673 0.658 0.635 0.648 0.655 0.660 0.660 0.655 14 0.636 0.661 0.658 0.645 0.661 0.665 0.662 0.670 0.662 15 0.650 0.630 0.658 0.646 0.655 0.662 0.659 0.653 0.660 16 0.635 0.643 0.647 0.653 0.650 0.659 0.659 0.660 0.640 17 0.642 0.635 0.643 0.661 0.646 0.656 0.660 0.655 0.670 18 0.653 0.653 0.659 0.648 0.659 0.654 0.650 0.662 0.673 19 0.635 0.646 0.644 0.643 0.655 0.664 0.658 0.649 0.651 20 0.646 0.662 0.660 0.657 0.667 0.663 0.651 0.658 0.666 n 12 13 14 15 16 17 18 19 20 12 0.646 13 0.657 0.671 14 0.664 0.656 0.675 15 0.666 0.670 0.663 0.663 16 0.668 0.662 0.659 0.661 0.663 17 0.658 0.654 0.654 0.661 0.664 0.661 18 0.671 0.657 0.657 0.657 0.647 0.655 0.667 19 0.662 0.669 0.651 0.655 0.666 0.654 0.658 0.662 20 0.658 0.659 0.656 0.664 0.662 0.651 0.658 0.672 0.656 Table B.7 Simulated critical values at percentile 97.5 for unbalanced samples without ties. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. The joint number of sample entries must be at least 8 for percentile 97.5 to be defined properly. 34 International Econometric Review (IER) n 3 5 4 5 0.745 0.727 6 7 8 9 10 11 6 0.707 0.71 0.716 0.722 7 0.725 0.684 0.708 0.728 0.745 8 0.739 0.714 0.702 0.733 0.729 0.719 9 0.694 0.74 0.737 0.738 0.74 0.743 0.746 10 0.709 0.718 0.73 0.742 0.745 0.738 0.747 0.76 11 0.675 0.74 0.725 0.746 0.766 0.734 0.73 0.728 0.727 12 0.689 0.722 0.72 0.722 0.726 0.73 0.756 0.743 0.744 13 0.701 0.74 0.745 0.727 0.738 0.749 0.753 0.75 0.743 14 0.711 0.724 0.74 0.732 0.727 0.745 0.755 0.759 0.754 15 0.692 0.711 0.736 0.736 0.739 0.742 0.747 0.746 0.74 16 0.702 0.727 0.715 0.718 0.749 0.74 0.717 0.744 0.74 17 0.705 0.714 0.728 0.743 0.72 0.737 0.737 0.738 0.767 18 0.713 0.729 0.725 0.727 0.737 0.735 0.726 0.758 0.76 19 0.696 0.718 0.723 0.731 0.746 0.749 0.73 0.727 0.739 20 0.7 0.73 0.74 0.734 0.755 0.747 0.734 0.762 0.751 n 12 13 14 15 16 17 18 19 20 12 0.731 13 0.737 0.762 14 0.741 0.743 0.756 15 0.757 0.773 0.743 0.736 16 0.75 0.753 0.734 0.753 0.74 17 0.741 0.743 0.736 0.747 0.76 0.741 18 0.758 0.751 0.746 0.75 0.738 0.744 0.75 19 0.737 0.759 0.736 0.747 0.756 0.747 0.747 0.764 20 0.745 0.734 0.748 0.742 0.743 0.739 0.737 0.756 0.743 Table B.8 Simulated critical values at percentile 99 for unbalanced samples without ties. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. The joint number of sample entries must be at least 9 for percentile 99 to be defined properly. 35 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… B.2. k = 3, ..., 15, balanced samples, no ties n 3 4 5 6 7 8 9 10 50 100 3 0.720 0.706 0.706 0.704 0.710 0.696 0.703 0.705 0.692 0.701 4 0.841 0.840 0.843 0.843 0.840 0.841 0.838 0.837 0.839 0.841 5 0.948 0.951 0.951 0.949 0.953 0.952 0.950 0.953 0.950 0.956 6 1.040 1.045 1.049 1.049 1.048 1.045 1.047 1.052 1.045 1.048 7 1.127 1.133 1.133 1.138 1.135 1.131 1.134 1.138 1.136 1.141 8 1.202 1.214 1.214 1.213 1.215 1.208 1.211 1.217 1.212 1.220 9 1.276 1.286 1.289 1.285 1.288 1.285 1.282 1.288 1.287 1.292 10 1.343 1.354 1.357 1.351 1.355 1.352 1.353 1.357 1.357 1.363 11 1.406 1.418 1.421 1.416 1.422 1.418 1.420 1.421 1.428 1.423 12 1.469 1.478 1.483 1.479 1.482 1.481 1.483 1.484 1.491 1.485 13 1.528 1.538 1.539 1.538 1.540 1.540 1.544 1.545 1.549 1.541 14 1.583 1.592 1.595 1.594 1.596 1.598 1.598 1.598 1.604 1.601 K 15 1.635 1.643 1.648 1.650 1.648 1.650 1.652 1.653 1.660 1.656 Table B.9 Simulated values of the HWM index at percentile 90 under A1* – A2*. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. n 3 4 5 6 7 8 9 10 50 100 3 0.753 0.762 0.779 0.773 0.778 0.770 0.774 0.780 0.776 0.773 4 0.886 0.900 0.904 0.909 0.906 0.906 0.904 0.911 0.910 0.915 5 0.993 1.004 1.012 1.016 1.011 1.015 1.013 1.020 1.025 1.020 6 1.087 1.102 1.109 1.109 1.111 1.111 1.109 1.117 1.121 1.122 7 1.177 1.189 1.191 1.198 1.198 1.197 1.199 1.202 1.204 1.211 8 1.254 1.267 1.273 1.273 1.276 1.280 1.278 1.276 1.283 1.294 9 1.327 1.341 1.346 1.346 1.350 1.347 1.350 1.352 1.361 1.365 10 1.395 1.406 1.411 1.414 1.422 1.419 1.417 1.425 1.431 1.430 11 1.454 1.474 1.477 1.480 1.485 1.484 1.484 1.489 1.497 1.494 12 1.518 1.535 1.536 1.539 1.542 1.545 1.545 1.548 1.559 1.551 13 1.576 1.590 1.597 1.598 1.601 1.603 1.603 1.609 1.618 1.611 14 1.631 1.644 1.654 1.655 1.655 1.661 1.660 1.665 1.671 1.667 K 15 1.684 1.699 1.706 1.712 1.709 1.713 1.713 1.719 1.726 1.720 Table B.10 Simulated values of the HWM index at percentile 95 under A1*– A2*. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. 36 International Econometric Review (IER) n 3 4 5 6 7 8 9 10 50 100 3 0.805 0.820 0.834 0.831 0.841 0.832 0.831 0.840 0.848 0.844 4 0.926 0.947 0.961 0.964 0.965 0.957 0.961 0.967 0.978 0.978 5 1.035 1.052 1.066 1.074 1.068 1.072 1.072 1.083 1.093 1.087 6 1.127 1.149 1.159 1.168 1.166 1.172 1.170 1.175 1.184 1.189 7 1.214 1.236 1.241 1.252 1.260 1.254 1.255 1.257 1.269 1.274 8 1.294 1.311 1.328 1.327 1.330 1.338 1.334 1.336 1.346 1.356 9 1.366 1.385 1.392 1.398 1.405 1.405 1.408 1.410 1.419 1.425 10 1.434 1.455 1.457 1.473 1.476 1.472 1.473 1.480 1.489 1.487 11 1.500 1.521 1.526 1.533 1.540 1.543 1.534 1.544 1.561 1.549 12 1.557 1.580 1.586 1.594 1.603 1.605 1.597 1.604 1.621 1.609 13 1.617 1.637 1.647 1.651 1.658 1.665 1.658 1.663 1.681 1.670 14 1.671 1.692 1.706 1.707 1.707 1.718 1.715 1.716 1.740 1.730 K 15 1.727 1.747 1.755 1.764 1.761 1.772 1.771 1.773 1.790 1.785 Table B.11 Simulated values of the HWM index at percentile 97.5 under A1*–A2*. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. n 3 4 5 6 7 8 9 10 50 100 3 0.839 0.874 0.888 0.896 0.909 0.900 0.906 0.917 0.919 0.930 4 0.968 1.008 1.013 1.024 1.033 1.027 1.037 1.042 1.057 1.055 5 1.082 1.111 1.125 1.135 1.142 1.144 1.138 1.150 1.172 1.169 6 1.174 1.208 1.212 1.225 1.233 1.247 1.244 1.246 1.256 1.254 7 1.259 1.287 1.301 1.310 1.318 1.327 1.329 1.327 1.341 1.341 8 1.333 1.371 1.382 1.394 1.392 1.393 1.406 1.394 1.416 1.422 9 1.412 1.434 1.454 1.464 1.470 1.479 1.470 1.472 1.494 1.490 10 1.477 1.505 1.522 1.532 1.534 1.543 1.532 1.551 1.557 1.547 11 1.545 1.574 1.585 1.600 1.599 1.603 1.600 1.612 1.636 1.623 12 1.609 1.628 1.645 1.658 1.660 1.664 1.667 1.670 1.700 1.686 13 1.665 1.685 1.704 1.714 1.719 1.724 1.728 1.727 1.757 1.740 14 1.721 1.748 1.758 1.771 1.767 1.784 1.784 1.782 1.805 1.805 K 15 1.783 1.801 1.814 1.821 1.822 1.836 1.836 1.833 1.859 1.860 Table B.12 Simulated values of the HWM index at percentile 99 under A1* – A2*. Notes: The underlying distribution consists of 10,000 independent HWM index values that are computed for samples that are drawn from a standard normal distribution. REFERENCES Anderson, T.W. and D.A. Darling (1952). Asymptotic theory of certain goodness of fit criteria based on stochastic processes. Annals of Mathematical Statistics, 23, 193-212. Cramér, H. (1928). On the composition of elementary errors II: statistical applications. Skandinavisk Aktuarietidskrift, 11, 141-180. 37 Hinloopen, Wagenvoort and Marrewijk- A k-sample homogeneity test:… Dufour, J.M. (1995). Monte Carlo tests with nuisance parameters: a general approach to finite-sample inference and non-standard asymptotics in econometrics. Technical report C.R.D.E., Université de Montreal. Dufour, J.M. and J.F. Kiviet (1998). Exact inference methods for first-order autoregressive distributed lags models. Econometrica, 66, 79-104. Fisz, M. (1960). On a result by M. Rosenblatt concerning the von Mises-Smirnov test. Annals of Mathematical Statistics, 31, 427-429. Girling, A.J. (2000). Rank statistics expressible as integrals under PP- plots and receiver operating characteristics curves. Journal of the Royal Statistical Society B, 62, 367-382. Hinloopen, J. (1997). Research and development, product differentiation, and robust estimation. Ph.D.-thesis, European University Institute. Hinloopen, J. and R.J.L.M. Wagenvoort (2010). Identifying All Distinct Sample P-P Plots, with an Application to the exact Finite Sample Distribution of the L1–FCvM Test Statistic. Tinbergen Institute Discussion Paper, TI 2010-083/1. Holmgren, E.B. (1995). The p-p plot as a method of comparing treatment effects. Journal of the American Statistical Society, 90, 360-365. IMF (2011). International Financial Statistics. http://www.imf.org (accessed March 31, 2012). Johnson, B.McK. and T. Killeen (1983). An explicit formula for the C.D.F. of the L1 norm of the Brownian bridge. The Annals of Probability, 11, 807-808. Kiefer, J. (1959). K-sample analogues of the Kolmogorov-Smirnov and Cramer-v.Mises tests. Annals of Mathematical Statistics, 30, 420-447. Kolmogorov, A.N. (1933). Sulla determinizazione empirica delle leggi di probabilita. Giornale dell ’Istituto Italiano Attuari, 4, 1-11. Kuiper, N.H. (1960). Tests concerning random points on a circle. Koninklijke Nederlandse Akademie van Wetenschappen, The Netherlands. Rosenblatt, M. (1952). Limit theorems associated with variants of the von Mises statistic. Annals of Mathematical Statistics, 23, 617-623. Scholz, F.W. and M.A. Stephens (1987). K-sample Anderson-Darling tests. Journal of the American Statistical Association, 82, 918-924. Schmidt, F. and M. Trede (1995). A distribution free test for the two sample problem for general alternatives. Computational Statistics & Data Analysis, 20, 409-419. Smirnov, N.V. (1939). On the deviation of the empirical distribution function. Mathematicheskii Sbornik, 6, 3-26. 38 International Econometric Review (IER) Von Mises, R. (1931). Wahrscheinlichkeitsrechnung. Vienna: Deuticke. Wilk, M.B. and R. Gnanadesikan (1968). Probability plotting methods for the analysis of data. Biometrika, 55, 1-17. 39